Forecasting Realized Volatility: The Choice of Window Size

ABSTRACT Different window sizes may produce different empirical results. However, how to choose an ideal window size is still an open question. We investigate how the window size affects the predictive performance of volatility. The empirical results show that the loss function for volatility prediction takes on a U‐shape as the window size increases. This suggests that if the window size is chosen too large or too small, the loss function tends to be large and the model's predictive accuracy decreases. A window size of between 1000 and 2000 observations is ideal for various assets because it can produce relatively minimal forecast errors. From an asset allocation perspective, a mean–variance investor can obtain sizeable utility by using a model with a low loss function value for her portfolio. Moreover, the results are robust in a variety of settings.